Anthropic has opened a research preview of Claude Duet, a standalone app that runs a proactive agent on Claude Opus 5.5. Rather than waiting for a prompt, Duet monitors connected tools, sends a morning brief and works on larger tasks on its own cloud computer, asking before sensitive actions. Access is by waitlist for US Pro and Max subscribers.
What Anthropic says Duet is
Per Anthropic's product page, Duet is "a proactive agent that understands what needs your attention." The listed behaviours are:
- Morning brief. It summarizes items that need attention and drafts next steps, then stays quiet until something needs the user.
- Larger tasks. It takes on work such as setting up meetings or building dashboards and reports back when done.
- Shared context. It uses the user's memory, skills and apps from the Claude app, in one message thread.
- Finished outputs. It returns docs, sheets or decks for review.
- Business examples. Flagging inbox and calendar items, drafting replies, finding unpaid invoices and drafting reminders, preparing payroll for review, and getting contracts ready to send.
Named connectors include Gmail, Google Calendar, Slack, Shopify, Stripe, QuickBooks, Google Drive, HubSpot, Notion and Microsoft 365. It ships as an iOS and Android app and on the web at claude.ai.
AnthropicThe approval model
The notable design choice is where permission sits. Anthropic says Duet works on its own cloud computer, where approvals, logins and safety checks run, and that it asks before sensitive actions such as sending messages or paying. Users can see what it has done and is doing, and steer it.
That places Duet in a trust-versus-autonomy trade-off that other agent launches have also faced. The more an agent can do unattended, the more the quality of its approval rules matters. Our coverage of AI agents moving customer money at JPMorgan, Schwab and Block shows how financial firms are restricting agents to read-only access, and xAI's Grok Bot launch covers a comparable cloud-computer approach.
| Design question | What Anthropic states | What is not yet public |
|---|---|---|
| When does it act? | Proactively, on morning briefs and ongoing tasks | How often it checks, and what triggers a task |
| What needs approval? | Sensitive actions such as sending messages or paying | The full list of gated actions |
| Where does work run? | A cloud computer with approvals, logins and safety checks | Isolation details and data retention |
| What does it cost? | No separate price; uses plan limits | How much background checking consumes allowance |
| Where can it fail? | Anthropic says it may misread a thread or misjudge urgency | Measured error rates |
The limits Anthropic lists
Anthropic is direct about three limits: Duet may misread a thread, may misjudge urgency, and can miss something that arrives between check-ins. It also says background checks count against plan usage, and larger tasks need more model turns. For a proactive tool, that last point matters, because cost scales with how often it looks, not only with what it is asked to do.
Analysis: why proactivity is the next product axis
Three shifts explain why labs are moving here.
- From pull to push. A chat product is only used when someone opens it. A proactive agent creates a daily touchpoint, and a morning brief is the simplest form of it.
- From answers to deliverables. Returning finished docs, sheets or decks changes what is measured, from answer quality to task completion.
- From single model to connected system. With 6,000-plus connectors, the product's value depends as much on integration breadth as on the model.
This follows Anthropic's earlier direction, covered in Claude Cowork and chat merging into one app. The open question is reliability. A morning brief that misjudges urgency by a small margin is a minor annoyance; an agent preparing payroll for review is a place where small errors have consequences, which is why the review step is central.
For proactive agents, the model is only part of the product. The approval rules decide how much of the work a person can safely hand over.
For users choosing tools
Duet is a research preview limited to US subscribers on a waitlist, so most people cannot test it yet. Teams evaluating proactive assistants will want to look at how approval is enforced, how usage is metered, and how portable their context is between providers. Metir takes a model-agnostic approach, so teams can try agents and models from multiple labs while keeping their own workspace and history.
What to watch next
- Expansion beyond US Pro and Max. Team and Enterprise access would signal confidence in the approval model.
- Independent evaluations. Published error rates on urgency and drafting would turn the stated limits into measurements.
- Usage impact. How quickly background checks consume plan limits.
- Competing launches. Similar proactive agents from other labs would confirm this as a category.
Sources:
Image credits
- Hero and in-body photo: "Dario Amodei at TechCrunch Disrupt 2023 01.jpg" by TechCrunch, licensed CC BY 2.0, via Wikimedia Commons. Pulled 2026-10-22 and reviewed visually before use.